The closed loop does not know about the menstrual cycle

Design

decentralised observational study of donated real-world data, hierarchical Bayesian state-space modelling of insulin sensitivity from CGM time series

Population

77 menstruating adults with type 1 diabetes and regular cycles using automated insulin delivery, contributing 380 menstrual cycles

Primary outcome

glycaemic outcomes, insulin requirements and modelled insulin sensitivity by menstrual cycle phase

Effect

insulin sensitivity +2.6% early follicular (CrI 0.3 to 5.0) and -2.6% midluteal (CrI -5.2 to -0.1); luteal phase had highest insulin dose, highest mean glucose and least time in range; pattern in 84.7%

Seventy-seven menstruating adults with type 1 diabetes using automated insulin delivery donated real-world data covering 380 menstrual cycles — continuous glucose monitoring traces, insulin delivery, carbohydrate entries and cycle dates. A hierarchical Bayesian state-space model estimated insulin sensitivity from the glucose time series across cycle phases.

The pattern is consistent and modest. Total daily insulin dose and carbohydrate intake both peaked in the luteal phase, and that phase also carried higher mean glucose and less time in range. Model-derived insulin sensitivity was 2.6% higher in the early follicular phase (credible interval 0.3 to 5.0) and 2.6% lower in the midluteal phase (credible interval -5.2 to -0.1). The direction held for 84.7% of participants, but with substantial variation between individuals.

The point is not the size of the population effect, which is small. It is that automated insulin delivery algorithms adapt to yesterday, not to a cycle they know nothing about, and a patient whose time in range falls predictably for a week every month is currently being managed as though that were noise. Two things follow for clinic. Ask menstruating patients whether their control varies with their cycle — many have noticed and few have been asked. And when reviewing a download, look at whether the poor weeks recur at an interval, before concluding the settings or the adherence are wrong. This is an observational study of self-selected data donors, so it describes a real phenomenon without quantifying it for any individual.

• Ask menstruating patients directly whether glycaemic control varies across their cycle
• Look for a recurring monthly pattern in a download before adjusting settings or questioning adherence
• Expect the luteal phase to need more insulin, and support patients who already adjust for it
• The population effect is small (about 2.6%) and individual variation is large — help each patient identify their own pattern
• Self-selected data donors using automated insulin delivery are not representative of all patients with type 1 diabetes

The statistics, in plain English

A credible interval is the Bayesian equivalent of a confidence interval: the 0.3 to 5.0 range for early follicular sensitivity just excludes zero, so the effect is probably real but small and imprecisely estimated. Because 84.7% of participants followed the population direction, the average conceals people who vary far more and people who do not vary at all — which is why this is a reason to look at each patient's own data rather than to apply a correction factor. Data donated by self-selecting users of expensive technology represent an engaged and well-resourced group.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.